optimization levels for 6-STABLE build{kernel,world}

[Gary Kline]

On Thu, Sep 14, 2006 at 10:02:37AM +1000, Mark Andrews wrote:
> 
> > On Wed, Sep 13, 2006 at 12:17:15PM +0300, Stefan Lambrev wrote:
> > > Hello,
> > > 
> > > Oliver Fromme wrote:
> > > >Marc G. Fournier <freebsd at hub.org> wrote:
> > > > > What are ppl currently using for CFLAGS/COPTFLAGS in /etc/make.conf for
> >  
> > > > > building kernel/world?  I know awhile back it wasn't recommended to go 
> > > > > above -O2, for instance, but suspect that has changed ... ?
> > > >
> > > >The best optimization is probably to not override the
> > > >defaults at all, because they're already pretty optimal.
> > > >In fact, by overriding the defaults there's a good chance
> > > >to make things worse.  :-)
> > > >
> > > >The default CFLAGS are "-O2 -pipe -fno-strict-aliasing".
> > > >Anything above -O2 isn't supported, and using -O2 without
> > > >-fno-strict-aliasing also isn't supported (and will create
> > > >broken code for some programs).  A common mistake is to
> > > >specify CFLAGS="-O2 -pipe" and omit -fno-strict-aliasing.
> > > >That'll shot you in the foot sooner or later.
> > > >
> > > >Best regards
> > > >   Oliver
> > > >

	[[ ... ]]

> > 
> > 
> > 	A couple of things.  Will having gcc unroll loops have any
> > 	negative consequences?  (I can't imagine how:: but better 
> > 	informed than to have something crash inexplicability.)
> > 	With 6.X safe at -O2 and with -funroll-loops, that should be
> > 	a slight gain, right?  (It also will make an upgrade from 5.5 
> > 	to 6.[12] that much more rational.)
> > 
> > 	[Dumb] questions:: first, what does the compiler do with
> > 	"-fno-strict-aliasing"?  And is there any guess, any SWAG even,
> > 	on when FreeBSD will safe with -O3??
> 
> 	Lots of code is not strict-aliasing safe.  Gcc itself can't
> 	determine all the cases which a construct is not strict-aliasing
> 	safe so even after getting rid of all the warnings gcc
> 	produces you can't be sure your code is strict-aliasing
> 	safe.  Think of -fstrict-aliasing as optimisation without
> 	a saftey net.  If your code doesn't cast pointers you should
> 	be safe otherwise you need to be really, really, really
> 	careful when you turn this on.
>  

	Well, my own pointer code is pretty trustworthy.  It's also
	pretty simple-minded and  thoroughly tested, even tho it's
	just for my own use.  Other's code, dunno.  People who cut their 
	teeth on pointers and may be just a wee bit cavalier: *Bzzz*
	Point well taken.

-- 
   Gary Kline     kline at thought.org   www.thought.org     Public service Unix